A Digital Palaeographic Approach towards Writer Identification in the

Dead Sea Scrolls

Maruf A. Dhali1 , Sheng He1 , Mladen Popović2 , Eibert Tigchelaar3 and Lambert Schomaker1
1 Institute of Artificial Intelligence and Cognitive Engineering (ALICE), Faculty of Mathematics and Natural Sciences,
University of Groningen, PO Box 407, 9700 AK, Groningen, The Netherlands
2 Qumran Institute, Faculty of Theology and Religious Studies,

University of Groningen, PO Box 407, 9700 AK, Groningen, The Netherlands

3 KU Leuven, Faculty of Theology and Religious Studies, Leuven, Belgium

{m.a.dhali, s.he, m.popovic}@rug.nl, eibert.tigchelaar@kuleuven.be, l.r.b.schomaker@rug.nl

Keywords: Dead Sea Scrolls, Handwritten Document Analysis, Digital Palaeography, Writer Identification, Handwriting
Recognition, Pattern Recognition, Feature Representation, Machine Learning.

Abstract: To understand the historical context of an ancient manuscript, scholars rely on the prior knowledge of writer
and date of that document. In this paper, we study the Dead Sea Scrolls, a collection of ancient manuscripts
with immense historical, religious, and linguistic significance, which was discovered in the mid-20th century
near the Dead Sea. Most of the manuscripts of this collection have become digitally available only recently
and techniques from the pattern recognition field can be applied to revise existing hypotheses on the writers
and dates of these scrolls. This paper presents our ongoing work which aims to introduce digital palaeography
to the field and generate fresh empirical data by means of pattern recognition and artificial intelligence. Chal-
lenges in analyzing the Dead Sea Scrolls are highlighted by a pilot experiment identifying the writers using
several dedicated features. Finally, we discuss whether to use specifically-designed shape features for writer
identification or to use the Deep Learning methods on a relatively limited ancient manuscript collection which
is degraded over the course of time and is not labeled, as in the case of the Dead Sea Scrolls.

1 INTRODUCTION With regard to choosing the right methodology,

This paper is part of a pioneering project on the
Dead Sea Scrolls that is sponsored by the European
Research Council (EU Horizon 2020). This multi-
disciplinary project brings together the natural sci-
ences, artificial intelligence, and the humanities in or-
der to shed new light on ancient Jewish scribal culture
by investigating two aspects of the scrolls’ palaeog-
raphy: handwriting recognition (the typological de-
velopment of writing styles) and writer identification.
Recognizing the handwriting would solve the when,
which and where questions, and identifying the writer
would end up answering the who question. These
are the four most important perspectives (figure 1) in
the study of palaeography and book history (Stokes,
2015). The digitization of the Dead Sea Scrolls (DSS)
has opened the door for pattern recognition to be ap-
plied in answering those four questions (4-W). We
aim to bridge the gap between computational science
and traditional palaeography by solving the 4-W with
a potential impact on digital palaeography beyond
DSS studies.
optical character recognition (OCR) methods are not
sufficient for historical manuscripts. There are mod-
ern forms of neural networks (Deep Learning) hav-
ing exceptionally good results (LeCun et al., 2015) in
many aspects of pattern recognition including hand-
written document analysis. But these performances
can only be achieved in case of millions of training
examples, which are contrary to the number of doc-
uments in many historical manuscripts, especially in
the DSS.
Here, we will present preliminary results of writer
identification in the DSS using several hand-crafted
features. Although this gives us fast results with-
out lengthy training on the limited labelled data of
the DSS, they are certainly not the best results to
be expected. We consider the results as a baseline
measurement for later experiments. We suggest how
to improve the results by exploiting the power of
parameter-heavy machine learning methods using this
small dataset. In solving this, we make a three-fold
proposition of advanced statistical modelling, data-
augmentation, and the use of pre-trained networks.

693
Dhali, M., He, S., Popović, M., Tigchelaar, E. and Schomaker, L.
A Digital Palaeographic Approach towards Writer Identification in the Dead Sea Scrolls.
DOI: 10.5220/0006249706930702
In Proceedings of the 6th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2017), pages 693-702
ISBN: 978-989-758-222-6
Copyright
c 2017 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
1.1 Dead Sea Scrolls
The DSS are a collection of ancient damaged
manuscripts that were discovered in the mid-20th cen-
tury in the Judaean Desert, in between Jerusalem and
the Dead Sea. Most were written over a period of
almost four centuries (ca. 250 BC to ca. 135 AD)
(Tigchelaar, 2010; Popović, 2012; Popović, 2015) in
characters commonly known as the Hebrew alpha-
bet, which actually derives from the Aramaic script
(Yardeni, 2002). The manuscripts, written by many
different writers, some of whom may have written
multiple manuscripts, display a broad variety and de-
velopment of different styles of this Hebrew-Aramaic
script. The study of ancient handwriting provides
the chronological framework, but the typological se-
quence of writing styles has to date not been system-
atically assessed for the DSS.
This project carries out the first systematic assess-
ment of the palaeographic framework of the scrolls
by combining two approaches. First, we will con-
duct new radiocarbon (14 C) dating on a number of
physical samples of the scrolls, kindly provided to
us by the Israel Antiquities Authority (IAA). Sec-
ond, we will generate for the first time quantitative
data for palaeographic handwriting recognition by
means of Artificial Intelligence, using the Monk sys-
tem, designed by Schomaker’s research group at AL-
ICE (Schomaker, 2016; Van der Zant et al., 2008;
Bulacu and Schomaker, 2007). The challenging is-
sue of writer identification in the DSS has not been
systematically dealt with before. The tools of digital
palaeography enable new, significant steps forward.
In this paper, we focus on this second approach of
digital palaeography.
Who? - Writer identification
When? - Temporal alignment
Which? - Manuscript identification
Where? - Localization
Figure 1: The four interesting questions for handwritten
manuscript understanding (image from the DSS manuscript
PAM 43.754, source: Brill scans).
1.2 Challenges in Digital Palaeography
In order to achieve both goals, i.e., handwriting recog-
nition and writer identification, specific challenges at
several levels of Computer Vision and Artificial Intel-
ligence must be overcome. Initial analyses are needed
for the proper extraction of characters (foreground,
ink) from the background, which is mostly either an-
imal skin or papyrus in the case of the DSS. Several
694
image processing techniques need to be applied for
optimum results of segmentation. Starting with edge
detection, morphological operations, filling gaps and
then finding connected components help to automati-
cally segment the hand-written fragments. Then fur-
ther processing can localize and extract the charac-
ters. Due to the difference in the textures of papyrus
and animal-skin, individual measures must be taken
on their distinctive periodic structures.
We have explored different feature-extraction
techniques on the images of the DSS. Feature-
representation maps the raw pixel intensity into a dis-
criminant high-dimensional space (Mikolajczyk and
Schmid, 2005; Li et al., 2015) in order to capture
specific information of the characters which can be
processed by algorithms in computers. This step is
an important element in the field of computer vision
and pattern recognition. There have been a lot of ef-
forts to design discriminative and powerful features
(Li et al., 2015). Though the (Deep) Learning-based
feature representation may achieve better results in
many cases, hand-crafted features have several advan-
tages in the analysis of handwritten documents, espe-
cially for historical manuscripts. This is due to the
amount of data in historical manuscript collections,
which is usually not big enough to train deep neural
networks. In contrast, the ImageNet data set (Deng
et al., 2009) contains millions of samples for train-
ing the network. The challenge becomes even higher
when the total number of usable pages comes to a
count of hundreds in the DSS. To take the opportu-
nities offered by the Deep Learning methods, the as-
sociated challenges need to be overcome in order to
analyse the DSS.
2 DATA
2.1 Manuscript Images
We will use digital images of the DSS as our primary
data. There are various sources for digital images of
the DSS manuscripts. The source used in this study
is kindly provided to us by Brill Publishers (Lim and
Alexander, 1995). There are 2463 images in the Brill
collection with varied resolutions from 600 by 600
pixels to 2800 by 3400 pixels, approximately. An-
other source is the high-resolution multi-spectral im-
ages of the DSS kindly provided to us by the Israel
Antiquities Authority (IAA), which derive from their
Leon Levy Dead Sea Scrolls Digital Library project.
In this project the IAA produces multi-spectral im-
ages of scrolls fragments on both the recto and verso
in 28 exposures, creating a file of 56 monochrome ex-
 A Digital Palaeographic Approach towards Writer Identification in the Dead Sea Scrolls
posures per fragment. The system then generates a
57th file of a colour image that combines all visible
wavelengths. The resolution of the files is 1,215 pix-
els per inch at a 1:1 ratio, capturing approximately
4 gigabytes of information per fragment (Shor et al.,
2014).
The Brill images are single-layered grayscale im-
ages with 300 ppi (pixels per inch) on both axis. They
have shadows and reflection from external lighting.
Additionally the lighting throughout all the images is
not uniform. Among the images, the ones containing
several fragments are mostly not aligned in a horizon-
tal way for text reading. This poses the issue of rota-
tion variance in characters. Many of the images also
contain paper calibration strips for scale representa-
tion and contemporary hand-written numbers. The
digitization noise can also be noticed in many of the
images (see figure 2).
Figure 2: Two of the Brill images; PAM 40.456 (left) and
PAM 40.531 (right); the images show digitisation noise,
alignment issues with the small fragments, shadows near
the border and lighting problems.
The IAA images are clear, properly aligned and
free from the problem of lighting and shadow un-
like the Brill ones. Additionally the different expo-
sure bands of the IAA hold important underlying in-
formation regarding the fragments providing essential
attributes for the scrolls. For example, one particu-
lar band provides clear information on the ink (fore-
ground) whereas another one gives more details on
the underlying leather/papyrus (background) on the
retro side. Some bands are useful for the textual con-
tents and some other bands give a better understand-
ing on the textural properties of the scroll material.
Extraction of this useful information is possible on
both single images and multi-spectral-fused images.
As a whole, digital image data has provided a new
and broader perspective in the quantitative analysis
and processing of the scrolls.
The scope of the current paper is limited to the
images in the Brill collection, but in the near future
we expect to publish our results on the digital data
from the IAA’s Leon Levy Dead Sea Scrolls Digital
Library. The quality and the challenges of the Brill
images can not be seen as a set-back, rather it would
be a starting benchmark to the robustness of our work.
Additionally, the possibility of using the IAA
material will improve our results.
2.2 Ground Truth
Unlike many other historical manuscripts, the DSS
do not have a structured and complete dataset nor the
ground truths for testing. Before diving into any sort
of computer aided writer identification, the ground
truths must be there to analyse the results. To estab-
lish the ground truths, we need experts in the field
and also their proper access to the data. We have this
two-folded advantage in our group: first through the
presence of palaeographic experts and second through
the Monk system which is accessible through web
browsers. By integrating these, we started to label the
DSS image data for ground truths.
We have proposed two different methods for la-
belling. The first one is to detect the region of interest
in the DSS images. The second one is to create the
ground truth for character-labels. Both these tasks re-
quire manual labour from experts with palaeographic
background knowledge on image level to pixel level.
This is the bench-mark in identifying the writers and
aligning the temporal developments in script style
for the DSS. In this paper, we will only use the la-
belled regions of interest (we call them FragmROIs,
by shortening the term fragment region of interest)
in order to build algorithms to extract features (using
available methods) and identify (recognition) writers.
From the DSS images, FragmROIs were selected and
labelled by the palaeographic experts using the Monk
system (see figure 3). Those rectangular FragmROIs
could consist of the entire text on an image, or of only
a section of text selected from an image. Different
FragmROIs from one and the same manuscript were
labelled as stemming from one writer or scribe, unless
palaeographers distinguished two scribes as writers of
the manuscript.
Figure 3: Using the Monk system, the palaeographic expert
can select the region of interest (FragmROI), then put the
associated attributes (scribe, style, comment etc.) to pro-
duce an XML file, which will later be used as labelled data.
Currently the experts can only select a rectangular region,
but the provision of choosing a polygonal region of interest
will be added to the system in near future.
695
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
While labelling the writers, we have set up a provi-
sional naming rule starting with the name scribeAxxx,
where xxx are numerical values starting from 001.
Each of the human labelled new-writers will be al-
located with an individual value. The term A is put
before the numerical values in order to preserve the
tag of original labelling from the palaeographic ex-
perts. If at a later stage of our study, two of the
writers are found to be the same one according to
the system, then they can be referred to with the new
name of scribeBxxx having two child-node of format
scribeAxxx, preserving the original label.
The present pilot study is based on two distinct
set of writers. The first set is a limited sample of
323 FragmROIs labelled as having been written by
13 scribes, namely the scribes of 1QIsaa columns
1-27, 1QIsaa columns 28-54, 1QS, 1QSa, 1QSb,
1QM, 1QpHab columns 1-12, 1QpHab columns 12
end-13, 4Q53, 4Q175, 11Q5, 11Q19 columns 2-
5, 11Q20. We labelled them from scribeA001 to
scribeA013. Distinct manuscripts were labelled as
deriving from different writers, even though in sev-
eral of the manuscripts of the first set palaeographers
think that one and the same writer produced multi-
ple manuscripts (Tigchelaar, 2002). To incorporate
the palaeographic opinion, we then merged those 13
scribes into 7 scribes by introducing scribeBxxx se-
ries. Then we took the second set of 13 scribes
with a limited sample of 124 FragmROIs labelled
as scribeA014 to scribeA026 (the scribes of 4Q266,
4Q504, 11Q10, 1Q22, 4Q209, 4Q167, 4Q6, 4Q286,
4Q381, 4Q405, 4Q491, 4Q431, 4Q525). The main
difference between these two sets are the amount of
characters per scribe. The first set has a higher num-
ber of characters than the second set. Thus, for this
pilot project, we have 447 FragmROIs labelled as
20 distinguishable scribes according to palaeographic
opinion.
3 A PILOT PROJECT
The DSS image data has its own distinctive char-
acteristics compared to other historical manuscript
datasets. This data set has with quite a different ap-
pearance from, e.g. historical manuscripts such as the
Medieval Palaeographic Scale data set (Monk, 2016)
from a previous project (He et al., 2016) in three as-
pects: (1) the number of characters in fragments from
some documents can be as low as one; (2) the ink of
each character has been faded out over the course of
time, making it more difficult to observe and process;
(3) the large diversity and lack of uniformity among
text blocks, presenting a challenge for analysis. In
696
this section, we will present the methodology used in
our pilot project in writer identification to benchmark
our works in analysing the DSS.
3.1 Writer Identification
Identifying writers using computers has been done for
decades (Plamondon and Lorette, 1989), which is a
problem of recognizing the writer of a given docu-
ment based on handwriting styles. A number of dif-
ferent features have been proposed and studied for
writer identification on scripts from several languages
including Dutch (Bulacu and Schomaker, 2007), En-
glish (Schomaker and Bulacu, 2004), Indic (Adak
and Chaudhuri, 2015; Karunakara and Mallikarju-
naswamy, 2011), and Arabic (Bulacu et al., 2007).
In the case of the DSS, we will be identifying the
scribes behind the scrolls with Hebrew characters, and
a hand-crafted feature specially for these characters is
yet to be proposed and studied. Instead of design-
ing a new feature, we initially started working with
some of the existing textural-based and grapheme-
based features. Textural-based features are based on
the statistical information about slant and curvature
of the handwritten characters, and grapheme-based
features, inspired by the bag-of-words model, extract
local structures and then map them into a common
space (He and Schomaker, 2016). We briefly discuss
the preprocessing techniques and the features used in
this work in the following sections.
3.1.1 Preprocessing
As the feature extraction technique is applied on the
binarized images, first we pre-processed the DSS im-
ages. Binarizing the Dead Sea Scrolls images is quite
challenging, given their diverse intensity, similarity
between ink and background traces, and image qual-
ity. We first started with Sobel edge detection (So-
bel, 1990) and then removed the connected objects on
the border to get rid of the markings. Morphological
operation was then used followed by image thresh-
olding. We used the global Otsu threshold selection
method (Otsu, 1975) as it is efficient and parameter-
less (see figure 4).
3.1.2 Feature Representation
Previous studies showed that the textural-based
feature extraction methods perform better than
grapheme-based methods (He and Schomaker, 2016;
He and Schomaker, 2017). Additionally, a more pow-
erful approach was introduced by using the spatial co-
occurrence among features (Bulacu and Schomaker,
2007; Ito and Kubota, 2010; Qi et al., 2014). The
 A Digital Palaeographic Approach towards Writer Identification in the Dead Sea Scrolls
Figure 4: The left one is a FragmROI from Brill collection
(PAM 43.787A) and the right one is the binarized image
using the Otsu threshold selection method.
later idea has been extended in a previous work (He
and Schomaker, 2017) with the introduction of the
joint feature distribution principle (JFD principle). By
accommodating these facts, we used eight textural-
based methods (three of them following the JFD prin-
ciple) and one grapheme-based method.
Hinge. The Hinge feature is the joint probability dis-
tribution of orientations of the legs of two contour
fragments attached at a common-end pixel on the ink
contours (Bulacu and Schomaker, 2007). Figure 5
shows two examples of the Hinge kernel on contour
fragments with leg length l and the joint probability
of the two orientations, α and β (α < β), are quan-
tized into a 2D histogram. Empirically we have set
l = 7 and the number of bins of α and β is set to 23.
Finally, the dimension of the feature vector is 253.
C(F1 )=1.06
C(F1 )=0.53
α1
α2
β1 β2
C(F2 )=1.06 C(F2 )=1.06
Figure 5: The two figures show two contour fragments with
the same Hinge kernel (α1 =α2 and β1 =β2 ) but different
fragment curvature values C (Fc ).
Co-occurrence Hinge (CoHinge). The CoHinge fea-
ture is the joint distribution of Hinge kernel following
the JFD principle on two different points xi and x j
with Manhattan distance l (figure 7) on the contours
as equation 1.
 
CoHinge(xi , x j ) = Hinge(xi ), Hinge(x j ) (1)
Each Hinge kernel has two values α and β, and
therefore, the CoHinge kernel has four values [α(xi ),
β(xi ), α(x j ), β(x j )], which can be quantized into a
4D histogram. The Manhattan distance l is set to
7 based on our previous study (He and Schomaker,
2017). We set the number of bins of the angle to 10,
and finally the dimension of the CoHinge feature is
10 ∗ 10 ∗ 10 ∗ 10 = 10, 000.
∆n Hinge. The ∆n Hinge is a rotation-invariant tex-
ture feature (He and Schomaker, 2014), computed by
building a feature network with the differential oper-
ator between Hinge kernels as the kernel function K i :
( n−1 n−1
∆n α(xi ) = ∆ α(xi )−∆δl α(xi +δl)
n−1 n−1 (2)
∆n β(xi ) = ∆ β(xi )−∆δl β(xi +δl)
where (α, β) is the Hinge kernel and n is the order
of the differential operator. Although many different
features can be generated based on the feature net-
work with different n, we work with the ∆1 Hinge fea-
ture with a feature-dimension of 780.
Quadruple Hinge (QuadHinge). QuadHinge is a
powerful feature representation following the JFD
principle, which incorporates the curvature informa-
tion of the contour fragments in the Hinge kernel by
computing a fragment curvature measurement (FCM)
C (Fc ) for contour fragments (Benhamou, 2004).
Quill and QuillHinge. The Quill feature (Brink et al.,
2012) is the joint probability distribution p(α, w) of
the relation between ink direction α and the ink width
w characterizing the writing material properties. The
QuillHinge is an extension of the Quill and Hinge,
which is the probability of p(α, β, w), resulting in a
3D histogram. We use the same parameters of the
Quill and QuillHinge as the original paper (Brink
et al., 2012), and the dimensions of Quill and Quill-
Hinge are respectively 1600 and 31200.
Triple Chain Code. The triple chain code feature
(Siddiqi and Vincent, 2010) is based on the chain code
on a pixel of the writing contours, which is the one of
eight directions where the next pixel is on, denoted
from 1 to 8.
TCC(xi , xi+l , xi+2l ) = [CC(xi ), CC(xi+l ), CC(xi+2l )]
(3)
where CC(xi ) ∈ {1, 2, · · · , 8} is the chain code value
on position xi , and l is the Manhattan distance along
the writing contours. We take the same value of l = 7,
similar as the CoHinge feature. The feature dimen-
sion is 512.
Cloud Of Line Distribution (COLD). COLD is a
curvature-free feature designed with the fact that writ-
ing contours can be approximated by a set of line seg-
ments obtained by the sequential polygonization al-
gorithm (Siddiqi and Vincent, 2010) and the lengths
and orientations of these straight lines can capture the
handwriting styles. The high ordered curvature points
on the writing contours are obtained using the method
(Prasad et al., 2011), denoted by P = {pi (xi , yi ), i =
0, 1, 2, · · · , n}, where (xi , yi ) is the coordinate of the
point pi (see figure 6). The line segments can be
obtained between any pair of the dominant points
(pi , pi+k ), where k is the parameter which denotes the
697
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
distance on the dominant sequence P . Each line can
be measured by a pair (θ, ρ) in the polar coordinate
space, where θ is the line orientation and ρ is the line
length. All the lines in a given handwritten document
can form a distribution in the polar coordinate space
and can be quantized into a log-polar histogram in-
spired by the Shape Context (Belongie et al., 2002).
The features obtained with k = 1, 2, 3 in the log-polar
space with the radius 7 and the angular intervals 12
are concatenated into one feature vector with the di-
mension: 7 ∗ 12 ∗ 3 = 252.
(a) (b) (c)
k=1
(d) (e)
k=2
(f) (g)
Figure 6: Illustration of the process of the COLD construc-
tion on the Shin character: (a) The given binarized con-
nected component; (b) The contour extracted from the bina-
rized image (a); (c) Detected dominant points (red points);
(d) Line segments (red lines) obtained between pair domi-
nant points when k = 1; (e) The distribution of lines from
(d) in the polar coordinate space; (f) Line segments when
k = 2 (Note that some long lines are not shown in order to
make the figure more clear); (g) The distribution of lines
from (f) in the polar coordinate space.
Junction Features. Junclets (He et al., 2015), a
grapheme-based feature, is the stroke-length distribu-
tion in every directions from 0 to 2π around a refer-
ence point (see Figure 7) inside the ink trace. When
the center point lies on the junction points, such as
the fork points and high curvature points on the skele-
ton line of the ink strokes, the corresponding feature
is the junction feature, which contain the junction in-
formation around the joint point. We have taken the
stroke length distribution in 120 directions equidis-
tantly sampled from 0 to 2π and the feature dimension
of each junction is 120.
698
xi
l
m
xj
Figure 7: Left: Co-occurrence patterns on ink contours.
Right: An illustration of the stroke-length distribution on
a reference point (the blue point in the center). The green
rays are the partial length in each direction, and the yel-
low curve is the distribution of the partial length in the po-
lar space. The red line is the skeleton line of the stroke
ink. m is the maximum measurable stroke length (He and
Schomaker, 2017).
3.1.3 Identification Methodology
Writer identification is simply answering the who
question. For a query document QscribeA
scripti
, where
x
scripti is the script of the hand-written manuscript and
scribeAx is the writer which we want to identify, all
scripti
the documents in the database dssscribeA ∈ DSS scripti
i
are sorted according to the feature distance between
scripti scripti
QscribeA and dssscribeA to produce a hit-list where the
x i
writer of the top document is assigned to scribeAx .
Here scribeAi is the label of all the writers and for our
case scripti is a single script of Hebrew. The nearest
neighbour classification method is performed using
the leave-one-out (Brink et al., 2012; Siddiqi and Vin-
cent, 2010) strategy. We take the query document out
and sort the remaining documents according to their
distance function to an output hit-list. For the distance
function of the feature vectors, we have taken the χ2
(chi-squared) distance for its better performance (Bu-
lacu and Schomaker, 2007).
4 RESULTS
In this section we present the performance of writer
identification based on the features and methodology
explained in section 3.1. 447 FragmROIs were used
for the pilot test. In the first set we took 323 Fragm-
ROIs with 13 writers labelled from scribeA001 to
scribeA013 having 74, 33, 14, 13, 26, 37, 58, 3, 25,
24, 4, 10 and 2 FragmROIs respectively. The first set
consists of writers with a large number of characters
in their corresponding FragmROIs.
We first calculated the feature vectors for all the
FragmROIs. Then we performed the writer identi-
fication using the methodology explained in 3.1.3.
We produce the output hit-list of all the FragmROIs
sorted out in accordance with their distance to the in-
 A Digital Palaeographic Approach towards Writer Identification in the Dead Sea Scrolls

put FragmROI. The top-n performance is calculated Table 3: The top-1 and top-10 performance (in percent-
when the query FragmROI is recognized as the writer age) of writer identification for another 13 scribes from
of the FragmROI on the top n of the hit-list. For ex- scribeA014 to scribeA026, with limited text fragments.
ample, the top-10 hit-list signifies the overall percent- Feature Top-1 Top-10
age of finding the same writer as input within the first Hinge 61.90 92.06
ten candidates (shortest distanced) of the output hit- CoHinge 62.69 89.68
list. Similarly the top-1 means the top-most candidate ∆1 Hinge 43.65 85.71
in the output hit-list corresponds to the same writer QuadHinge 63.49 90.47
as the input. The performance of top-1 and top-10 Quill 48.38 89.51
hit-list for first set is presented in Table 1. Accord- QuillHinge 45.16 74.19
TripleChainCode 61.90 88.09
Table 1: The top-1 and top-10 performance (in percentage) COLD 58.87 88.71
of writer identification for 13 scribes from scribeA001 to
scribeA013.
Junclet 31.45 76.74

Feature Top-1 Top-10 Table 4: The top-1 and top-10 performance (in percentage)
Hinge 87.61 97.83 of writer identification for all 20 scribes.
CoHinge 81.11 95.97
Feature Top-1 Top-10
∆1 Hinge 79.87 94.73 Hinge 78.30 94.40
QuadHinge 89.47 96.59 CoHinge 79.19 89.93
Quill 80.80 93.80
∆1 Hinge 68.23 85.48
QuillHinge 76.78 89.78
QuadHinge 79.64 89.04
TripleChainCode 84.82 96.59
Quill 71.58 86.57
COLD 82.35 94.42
QuillHinge 69.57 82.10
Junclet 81.42 95.04
TripleChainCode 79.19 91.28
ing to the majority palaeographic opinion, scribeA001 COLD 76.95 88.37
and scribeA002 are the same scribe, and so also Junclet 40.71 72.93
scribeA003, 004, 005, 008, 009 and scribeA010, 012,
which are then labelled as scribeB001, B002 and
B003 respectively. The result is presented in Table 2
for these seven scribes. Then we took the second set
5 DISCUSSIONS
Table 2: The top-1 and top-10 performance (in percentage) 5.1 Performance Evaluation
of writer identification for 7 scribes: scribeB001, B002,
A006, A007, B003, A011 and A012 .
We presented the results of writer identification on a
Feature Top-1 Top-10 limited number of scribes. Of the shape based meth-
Hinge 92.26 98.76 ods, the QuadHinge performs the best in the top-1
CoHinge 93.80 97.52 hit-lists for three out of four cases (only for the case
∆1 Hinge 90.71 96.28 of seven scribes in Table 2, CoHinge performs better
QuadHinge 93.50 96.90 with a small difference of 0.30% than QuadHinge),
Quill 88.54 96.28 whereas the Hinge feature gives better result in all
QuillHinge 88.85 96.90 the top-10 hit-lists. The reason for this performance
TripleChainCode 92.26 98.14 can be deduced from the design criteria of the fea-
COLD 91.33 96.28 tures themselves. Hinge feature takes into account
Junclet 56.03 88.85 the joint probability distribution of the orientations of
legs of two contour fragments from a common end
of 13 scribes with 124 FragmROIs. The amount of pixel on ink contours, which proves to be a strong
text is lower in this set than in the first one. The identical property for individual scribes of these an-
result is shown in Table 3. Finally, we took all the cient manuscripts. Additionally, the incorporation of
scribes together for testing. Table 4 presents the result FCM to the Hinge feature following the JFD principle
of these 20 scribes together ( i.e., scribeB001, B002, gives the QuadHinge feature a boosted performance.
B003, A006, A007, A011, A013, A014 to A026). We The directional measurement of the ink-trace
briefly discuss the results and our propositions in the width makes the Quill feature, which is quite infor-
next section (Section 5). mative on quill-based medieval scripts, a weak can-

699
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
didate for the DSS. This is due to the uniformity of
the ink-trace in these documents coming from a prob-
ably fairly blunt tip of the ancient writing equipment.
Consequently, the QuillHinge fails to provide a higher
performance in this test set. The ∆1 Hinge has a lim-
ited performance, indicating that on Hebrew charac-
ters, loss of the angle with respect to the horizontal
removes too much of the writer-specific information.
The grapheme-based feature, Junclets, gives lower
performance than the cross-script writer identification
(He et al., 2015) due to the lower variability in the
stroke-length distribution in every direction around a
reference point inside the ink of the DSS’ Hebrew
characters.
5.2 Propositions
The challenges in analysing the DSS are unique
and unprecedented. Using the dedicated features
(in 3.1.2), we found fast results without lengthy train-
ing on the limited labelled data of the DSS. But
they are certainly not the best results to be expected.
Especially when the amount of data is small with
large variability (as in Table 3), the performance be-
comes lower. To overcome this situation, we need to
consider a pragmatic approach incorporating several
propositions.
1) Statistical modelling can be used in the case
of the DSS, where the sample size is low and there
are differences in the scholarly opinion of writers as
well. We can use the differences in writing attributes
of a set of different manuscripts to build a population
model. A writer model can be built using the query
manuscript. The classification is then carried out by
evaluating the similarity of a further manuscript sam-
ple with respect to the models. We can build our pro-
visional model, similar to the work of speaker identi-
fication (Leuzzi et al., 2016), as follows:
pb (d(Wi ,W j ))
Λ(d(Wi ,W j )) = (4)
pw (d(Wi ,W j ))
Here, d(Wi ,W j ) is the distance computed from Wi ,
the query writer to W j , the suspected writer. Λ de-
notes the likelihood ratio over d(Wi ,W j ). The distri-
bution of distances between the suspected writer and
the population is denoted by pb (d(Wi ,W j )), which
can be referred as the between-group distance among
the writers. pw (d(Wi ,W j )) is the distribution of dis-
tances taken within different instances of the sus-
pected writer (within-group distance). The collection
of statistical models (Fisher, 1925), analysis of vari-
ance (ANOVA), can be used to analyse the within-
group and between-group variances of the writers.
2) Another possibility is transfer learning (Long
and Wang, 2015). It starts with the use of pre-trained
700
networks on massive not-labelled handwriting collec-
tions. Such networks are trained to reconstruct im-
ages over (via) a very limited number of values (hid-
den units). After training, such a network implicitly
knows a lot about historical handwritings in general.
In a second stage, such a network is then applied to
the DSS, using those hidden unit vectors as feature
descriptors.
3) Data augmentation can be utilized in the pro-
cessing of the DSS. If there is a believable random
transformation of the DSS’ text patterns, i.e., one that
remains legible by humans, then for each natural sam-
ple of a character, a number of N derived random ver-
sions of it may be added to the training set, effectively
enlarging the amount of labelled data. Known already
in the nineties (Baird, 1992) this was later made pop-
ular in handwriting recognition later by the use of hid-
den Markov models (Varga and Bunke, 2003; Ha and
Bunke, 1997).
5.3 Conclusions
In this paper, we have introduced digital palaeogra-
phy of the DSS by presenting a pilot project, which
is part of a pioneering multi-disciplinary project that
brings together the natural sciences, artificial intel-
ligence, and the humanities. By introducing the
rule to establish ground-truths, we performed writer-
identification tests using dedicated features on provi-
sionally labelled data. The varying performance of
results for different sets of writers led us to the propo-
sitions of statistical modelling, transfer learning, and
data augmentation for this largely diverse collection
of manuscripts.
We consider the results of this paper as a baseline
measurement for our later experiments. We will com-
bine both the aspect of specifically-designed shape
features and the Deep Learning methods to produce
fresh empirical data for the study of the DSS. Addi-
tionally, we will conduct new radiocarbon (14 C) dat-
ing on a number of physical samples of the scrolls.
The outcome of 14 C dating will then be subjected to
Bayesian statistics methods in combination with the
results from temporal alignment using pattern recog-
nition to reach more accurate and precise dating of the
DSS.
ACKNOWLEDGEMENTS
The authors would like to thank Ruwan van der Iest
(research assistant for the ERC project at the Qum-
ran Institute) for his valuable inputs in labelling the
regions of interest through the Monk-system.
 A Digital Palaeographic Approach towards Writer Identification in the Dead Sea Scrolls
This work has been supported by an ERC Starting
Grant of the European Research Council (EU Hori-
zon 2020): The Hands that Wrote the Bible: Digital
Palaeography and Scribal Culture of the DSS (Hand-
sandBible # 640497). Additional support comes from
NWO (Netherlands Organisation for Scientific Re-
search) and FWO (the Research Foundation Flan-
ders): Models of Textual Communities and Digital
Palaeography of the DSS (# 326-25-001).
REFERENCES
Adak, C. and Chaudhuri, B. B. (2015). Writer identification
from offline isolated bangla characters and numerals.
In ICDAR, pages 486–490. IEEE.
Baird, H. S. (1992). Document image defect models. In
Structured Document Image Analysis, pages 546–556.
Springer.
Belongie, S., Malik, J., and Puzicha, J. (2002). Shape
matching and object recognition using shape contexts.
IEEE PAMI, 24(4):509–522.
Benhamou, S. (2004). How to reliably estimate the tortu-
osity of an animal’s path: straightness, sinuosity, or
fractal dimension? Theoretical Biology, 229(2).
Brink, A., Smit, J., Bulacu, M., and Schomaker, L. (2012).
Writer identification using directional ink-trace width
measurements. PR, 45(1):162–171.
Bulacu, M. and Schomaker, L. (2007). Text-independent
writer identification and verification using textural and
allographic features. IEEE PAMI, 29(4):701–717.
Bulacu, M., Schomaker, L., and Brink, A. (2007). Text-
independent writer identification and verification on
offline arabic handwriting. In ICDAR, volume 2,
pages 769–773. IEEE.
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-
Fei, L. (2009). ImageNet: A Large-Scale Hierarchical
Image Database. In CVPR.
Fisher, R. A. (1925). Statistical methods for research work-
ers. Genesis Publishing Pvt Ltd.
Ha, T. M. and Bunke, H. (1997). Off-line, handwritten
numeral recognition by perturbation method. IEEE
PAMI, 19(5):535–539.
He, S., Samara, P., Burgers, J., and Schomaker, L. (2016).
A multiple-label guided clustering algorithm for his-
torical document dating and localization. IEEE Trans-
actions on Image Processing, 25(11):5252–5265.
He, S. and Schomaker, L. (2014). Delta-n hinge: Rotation-
invariant features for writer identification. In ICPR,
pages 2023–2028.
He, S. and Schomaker, L. (2016). Writer identification us-
ing curvature-free features. PR.
He, S. and Schomaker, L. (2017). Beyond ocr: Multi-
faceted understanding of handwritten document char-
acteristics. PR, 63:321–333.
He, S., Wiering, M., and Schomaker, L. (2015). Junction
detection in handwritten documents and its applica-
tion to writer identification. PR, 48(12):4036–4048.
Ito, S. and Kubota, S. (2010). Object classification us-
ing heterogeneous co-occurrence features. In Euro-
pean Conference on Computer Vision, pages 701–714.
Springer.
Karunakara, K. and Mallikarjunaswamy, B. (2011). Writer
identification based on offline handwritten document
images in kannada language using empirical mode de-
composition method. Writer, 30(6).
LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learn-
ing. Nature, 521(7553):436–444.
Leuzzi, F., Tessitore, G., Delfino, S., Fusco, C., Gneo, M.,
and Zambonini, G. (2016). A statistical approach to
speaker identification in forensic phonetics field.
Li, Y., Wang, S., Tian, Q., and Ding, X. (2015). Feature
representation for statistical-learning-based object de-
tection: A review. PR, 48(11):3542–3559.
Lim, T. and Alexander, P. (1995). Volume 1. In The Dead
Sea Scrolls Electronic Library. Brill.
Long, M. and Wang, J. (2015). Learning transferable
features with deep adaptation networks. CoRR,
abs/1502.02791, 1:2.
Mikolajczyk, K. and Schmid, C. (2005). A perfor-
mance evaluation of local descriptors. IEEE PAMI,
27(10):1615–1630.
Monk (2016). Medieval palaeographic scale data set (online
collection).
Otsu, N. (1975). A threshold selection method from gray-
level histograms. Automatica, 11(285-296):23–27.
Plamondon, R. and Lorette, G. (1989). Automatic signature
verification and writer identification - the state of the
art. Pattern recognition, 22(2):107–131.
Popović, M. (2012). Qumran as scroll storehouse in times
of crisis? a comparative perspective on judaean desert
manuscript collections 1. Journal for the Study of Ju-
daism, 43(4-5):551–594.
Popović, M. (2015). The ancient ‘library’ of qumran be-
tween urban and rural culture. In The Dead Sea Scrolls
at Qumran and the Concept of a Library, pages 155–
167. Brill.
Prasad, D. K., Quek, C., Leung, M. K., and Cho, S.-Y.
(2011). A parameter independent line fitting method.
In ACPR, pages 441–445.
Qi, X., Xiao, R., Li, C.-G., Qiao, Y., Guo, J., and Tang,
X. (2014). Pairwise rotation invariant co-occurrence
local binary pattern. IEEE PAMI, 36(11):2199–2213.
Schomaker, L. (2016). Design considerations for a large-
scale image-based text search engine in historical
manuscript collections. it-Information Technology,
58(2):80–88.
Schomaker, L. and Bulacu, M. (2004). Automatic writer
identification using connected-component contours
and edge-based features of uppercase western script.
IEEE PAMI, 26(6):787–798.
Shor, P., Manfredi, M., Bearman, G. H., Marengo, E.,
Boydston, K., and Christens-Barry, W. A. (2014). The
leon levy dead sea scrolls digital library: The digitiza-
tion project of the dead sea scrolls. Journal of East-
ern Mediterranean Archaeology and Heritage Studies,
2(2):71–89.
701
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods

Siddiqi, I. and Vincent, N. (2010). Text independent

writer recognition using redundant writing patterns
with contour-based orientation and curvature features.
PR, 43(11):3853–3865.
Sobel, I. (1990). An isotropic 3× 3 image gradient operator.
Machine Vision for three-demensional Sciences.
Stokes, P. A. (2015). Digital approaches to paleography
and book history: some challenges, present and future.
Frontiers in Digital Humanities, 2:5.
Tigchelaar, E. (2002). In search of the scribe of 1qs. In
Emanuel, pages 339–352. Brill.
Tigchelaar, E. (2010). Dead sea scrolls. In The Eerdmans
Dictionary of Early Judaism, pages 163–180. Eerd-
mans.
Van der Zant, T., Schomaker, L., and Haak, K. (2008).
Handwritten-word spotting using biologically in-
spired features. IEEE PAMI, 30(11):1945–1957.
Varga, T. and Bunke, H. (2003). Effects of training set
expansion in handwriting recognition using synthetic
data. In Proc. 11th Conf. of the Int. Graphonomics
Society, pages 200–203. Citeseer.
Yardeni, A. (2002). The book of Hebrew script: his-
tory, palaeography, script styles, calligraphy & de-
sign. Oak Knoll Pr.

702